The Gift of Living Memory

ChatGPT spreads this gift across four places, and most of the skill is knowing which one a given line belongs in. Custom Instructions are for what is true of you in every chat - your town, your trade, the language and the length you want - so the About me and How to answer headings go there once and never into a prompt again. A Project holds one area of work with its own instructions and its own uploaded files, and those instructions apply only inside it, which is what you want when your shop's rules should not bend your child's homework answers. A GPT is the last step: once a Project's instructions have stopped changing, save them as a named assistant with its own conversation starters and the job becomes a button. Memory is the odd one out - it fills itself from what you happen to say, so it is the part you audit rather than the part you write.

From the book Use Custom Instructions for who you are and how you like answers everywhere, a Project for one area of work, and a GPT for a job you repeat often. Check what Memory has saved from time to time and delete anything outdated or private.

• About me and How to answer belong in Custom Instructions - written once.
• Give each area its own Project so rules do not bleed across.
• Promote a settled Project's instructions into a named GPT.
• Audit saved Memory monthly; delete stale facts and anything private.
• Never put customer, student or patient records in a shared GPT.

**Part one - Settings, Custom Instructions.** This is true of me in every chat, so it lives here:

> I am an ASHA worker in a village near Nashik. I speak with mothers, pregnant women and elders. Answer in simple spoken Marathi first, then English below it. Short sentences. Under 100 words unless I ask for more. Steps as a numbered list. Ask me one question whenever my request is unclear.

**Part two - a Project called “Village Health Messages”.** Put this in the Project instructions:

### Always
- Follow official government health guidance - National Health Mission, the Maharashtra health department, or my ANM's own instruction - and name which one you followed.
- Close anything medical with: ask at the PHC or your ANM.

### Never
- Never name a medicine or give a dose.
- Never ask me for a patient's name, phone number or record. I will type [name] and you will leave it as [name].

### Files in this Project
Only the monthly health-day calendar, with no names in it.

**Part three - the first real prompt, in a new chat inside that Project:**

Write the WhatsApp reminder for the vaccination day on [date] at the anganwadi. Before the message, give me one line saying which of my Always rules shaped it, so I can tell whether the Project instructions are actually working.

I want to build this once as a GPT called **Home Budget Helper**, so give me the finished text to paste into the builder's instruction box.

It must say this, under four headings:

**About me** - I run the house for a family of five in Kozhikode. The monthly budget is about ₹[amount]. I type amounts and nothing else.

**How to answer** - The table first, every time: Category | Planned | Spent | Difference, with the Malayalam and the English name in the same cell. Totals at the foot. Then two lines of comment and nothing more.

**Always** - Show the difference in rupees and as a percentage of the planned figure. Finish with one practical saving idea for a Kerala kitchen or an electricity bill, not general advice.

**Never** - Never ask for bank, card or UPI details or an account number. If I paste anything containing them, tell me to remove it and stop there.

Also write the three conversation starters for the GPT - the three things I will type each month.

Then, before I save it, test it: pretend I typed groceries ₹9,200, electricity ₹1,450, school fees ₹3,000, gas ₹950, other ₹2,100, and show me exactly what this GPT would answer.

**Custom Instructions first**, because this much is true of me outside this one course as well:

> Second-year computer science student in Toronto; I did my schooling in Nagpur. I learnt to program in Python and I read it far faster than Java. Explain things the way a senior student explains them across a hostel desk, not the way a textbook sets them out.

**Then a Project called “Data Structures”**, with the syllabus, the weekly lecture slides and the lab sheets in its files. Its Project instructions:

- **How to answer** - Open with the idea in three lines, then the full explanation, then one small snippet I can run. Anything long goes into Canvas, so that I can annotate it instead of scrolling back for it.
- **Always** - Begin by naming the uploaded file and the week you are answering from. Mark any sentence that does not come out of my files with “(outside the course)”.
- **Never** - Never produce a submittable answer to anything on the lab sheets. Hints, the structure to reach for and the first step, nothing further. If I push for more, quote this line back at me instead of obeying it.
- **Check on me** - Finish with one question I ought to be able to answer. If I get it wrong, do not hand me the answer; name the part of my own notes to read again.

First question in the Project: hash tables, from the week 4 material.